A Smile the Eyes Agree With
Almost every tool in this category fails the same way and for the same reason. A smile is not a mouth shape. A genuine one recruits a muscle that circles the eye, which narrows the opening, lifts the cheek, and creases the skin at the outer corner — and that happens before you notice the mouth at all. Change the mouth alone and a viewer gets a face making two contradictory statements, which is precisely what uncanny means.
Use this without the search next time. Prathom Workbench puts Prathom's tools in your toolbar.
Add to Chrome — freeDrop an image here, or click to browse
Up to 25 MB. Uploaded temporarily, deleted after processing.
What it does
- Cheeks, eyes, and mouth moved together
- Eye corner lines allowed to appear
- Identity and teeth held in place
- Moderate denoise, the whole face changes
How to use AI Smile Enhancer
- 1
Consider whether the photo should change
Altering someone's expression alters what the picture says happened. For your own portrait that is fine. For a photograph of an event it is a different act.
- 2
Run the pass
The expression is rebuilt as a whole — cheeks lift, eyes narrow slightly, the lines beside them appear — so the two halves of the face agree.
- 3
Cover the mouth and look at the eyes
Hold a finger over the mouth in the result. If the eyes alone do not read as amused, the smile will not convince anyone.
How it works
Your portrait is scaled to about one and a fifth megapixels and repainted at a denoise noticeably higher than the other portrait pages, with a fresh seed each run. The instruction describes a whole-face change rather than a mouth change: lift the cheeks, narrow the eyes slightly, allow the lines beside them to appear, and make the mouth and the eyes agree.
The negative prompt leads with the failure this page exists to avoid — a mouth only smile — followed by staring eyes, an uncanny expression, a forced grin, and changed teeth.
The eyes go first
Watch someone react to something funny and the order is fixed. The eyes narrow, the cheeks come up, and the mouth follows. The muscle responsible circles the eye socket and is largely outside voluntary control, which is why people find it hard to fake a convincing smile on demand and easy to spot one that has been faked.
That gives a viewer a reliable test they apply without effort: does the upper half of the face confirm what the lower half is claiming? A photograph where it does not is not read as a slightly imperfect smile. It is read as a person who is not actually happy, or as an image that has been altered, and both are worse than the neutral expression you started with.
So the useful design rule is that a smile tool is really an eye tool with a mouth attached, and this workflow is written that way.
The lines you spent the last page removing
There is a direct contradiction between this page and the wrinkle page and it is worth naming.
The wrinkle page softens the fan of lines beside the eye. This page needs that fan to appear, because it is the visible evidence of the muscle that makes a smile real.
Both are correct in their own context, and running them in sequence is a mistake. If you smooth the eye corners and then add a smile, the pass has to reinstate exactly what was just taken out, and what it reinstates is generated rather than yours. Pick one operation for a given photograph.
If you want both a smile and softer skin, do the expression first and go gently afterwards.
Whose photograph is it
The technical part of this page is straightforward. The rest is not, and it is short.
A portrait of yourself is yours to adjust. A photograph of a group at an event records something that happened, and changing how a person in it looked changes what it records. Family photographs get shared, reused, and eventually treated as evidence of what a day was like, and someone who was actually miserable that afternoon may not want a version of themselves enjoying it in circulation.
Ask. It takes less time than the run.
Publication gate
This page ships once the workflow has been run against a neutral headshot, an already forced smile, a portrait wearing glasses, and a subject at a three quarter angle, confirming in each case that the eye region changes together with the mouth and that the teeth are unaltered.
Examples
Neutral headshot
The intended case. A relaxed face has room to move in every region, which is what a whole-face expression change needs.
Forced smile
Harder than it looks. The mouth is already at its limit, so the work is entirely around the eyes, and that is the region people scrutinise.
Frequently asked questions
Why do edited smiles look fake?
Because they usually are only a mouth. A genuine smile contracts the muscle ringing the eye, which narrows the aperture, raises the cheek, and creases the outer corner. Viewers check this automatically without knowing they do, so a wide mouth beneath unchanged eyes reads as a person pretending rather than a person smiling.
Does it change my teeth?
It is instructed not to, and changed teeth appear in the negative prompt. This matters because teeth are individual — spacing, shape, and slight irregularity are part of recognizing someone — and a set of even generated teeth is one of the fastest ways to make a portrait read as synthetic.
Is it wrong to add a smile to a photo?
It depends entirely on what the photograph is for. Adjusting your own headshot is unremarkable. Changing how someone looked at a wedding, a protest, or a funeral edits the record of an event, and the person in the picture generally has an opinion about that worth asking for.
Why does it need a higher denoise than the other portrait pages?
Because it is the only one making a structural change. Retouching and wrinkle work adjust the surface of a face that stays where it is. An expression moves cheeks, eyelids, lips, and the skin between them together, so there is genuinely more to redraw and the identity guard has to work harder.
Is the AI Smile Enhancer tool free, and do I need an account?
Free, and deliberately account-free. Prathom has no login anywhere on the site, so there is no usage counter attached to you and no upgrade prompt waiting at the end of the job. The practical limits are the workflow's own: an image up to 25 MB going in, one queued GPU job at a time, and a result that stays available for thirty minutes.